Quality-Diversity Automatic Summarization based on Determinantal Point Processes

Information age requires intelligent tools like automatic abstract. We present a unique perspective from unsupervised data driven technology for extractive summaries. Inspired by a novel sampling method based on DPPs, we explore the latent cohesion among sentences through computing Word Mover's Distance (WMD) contrasted to traditional similarity metrics, and build condensing summaries with high quality and low redundancy by selecting diverse sentences from raw documents. We illustrate the advantages of our competing system primarily on the dataset from the 4^th Computational Linguistics Scientific Document Summarization Shared Task and MultiLing 2015 Task, and the performance is evaluated through F-measure from ROUGE family.